Zuteilung anfragen Kontakt English简体中文

WissenschaftTechnische Berichte

JM-TR-2019-004 Technischer Bericht Bibliothek v2.0

Library v2.0: a versioned multiple-sequence-alignment architecture for fragmentary ancient peptides, with a post-mortem of fragment loss during automated bracketing (builds v1.1–v1.4)

Dr. Anders Kjeldsen 1,* Dr. Hao Lin 1 Dr. Rahul Venkataraman 1 Dr. Yuting Zhou 1

1 Jurassic Meat Shenzhen Research Center, Shenzhen, China
* Korrespondierende Autorin / korrespondierender Autor: Dr. Anders Kjeldsen · regulatory@jurassicmeat.com

Inhalt
  1. Zusammenfassung
  2. 1. Summary
  3. 2. Introduction
  4. 3. Materials and Methods
  5. 4. Results
  6. 5. Discussion
  7. 6. Limitations and known issues
  8. 7. Changes in Library v2.0
  9. 8. Data availability
  10. 9. Acknowledgements
  11. 10. References
  12. Supplementary Table S1

Zusammenfassung

We describe Library v2.0, a re-architecture of the Jurassic Meat Peptide Library in which fossil-derived peptide fragments and extant reference sequences are held as versioned, locus-partitioned multiple sequence alignments (MSAs) keyed to a fixed reference coordinate system, replacing the unaligned per-specimen peptide files used since v1.1. The release increases theropod-assigned fragments from 389 to 1,284, unique residues covered from 3,050 to 7,920, and extant reference taxa from 19 to 34; median posterior placement probability for theropod-assigned fragments rises from 0.47 to 0.63. During migration, a reconciliation check identified a defect introduced into the automated bracketing module in August 2014: fragments for which an exact anchor could not be found in the reference profile were omitted from bracketing without error and without appearing in the unmapped-fragment log. In the final v1.4 build, 141 of 389 fragments (36.2%) were affected, disproportionately those shorter than eight residues and those carrying annotated deamidation. The defect was present for 51 months before detection. Retrospective re-bracketing of all 14 reconstruction targets altered 37 inferred residues across nine targets. Two research cell lines had been constructed from affected targets; neither progressed beyond research use. The impact assessment concluded that there was no impact on released material. We report corrective actions, including mandatory input reconciliation and logged exclusion for every build.

Schlagwörterpalaeoproteomicsmultiple sequence alignmentphylogenetic placementphylogenetic bracketingdata versioningpost-incident reviewcollagenenamel proteins

JM-TR-2019-004 · Library v2.0 · 4 June 2019 Dr. Anders Kjeldsen*, Dr. Hao Lin, Dr. Rahul Venkataraman, Dr. Yuting Zhou Shenzhen Research Center, Jurassic Meat, Nanshan District, Shenzhen *Corresponding author

Public summary of internal report (full internal version: 41 pages).

1. Summary

  • Library v2.0 replaces unaligned per-specimen peptide files with versioned, locus-partitioned multiple sequence alignments in a fixed reference coordinate system. Every build is now reproducible from a single content hash.
  • The library grows to 1,284 theropod-assigned fragments (v1.4: 389), 7,920 unique residues (v1.4: 3,050) and 34 extant reference taxa (v1.4: 19). Median placement posterior for theropod-assigned fragments is 0.63 (v1.4: 0.47).
  • A defect in the bracketing scripts (in service August 2014 – November 2018) caused fragments without an exact reference anchor to be silently omitted from bracketing. In the v1.4 build, 141 of 389 fragments were affected.
  • Retrospective re-bracketing changed 37 inferred residues across 9 of 14 reconstruction targets. The impact assessment (IA-2018-07) concluded that there was no impact on released material.
  • Every build now fails if the number of fragments entering bracketing differs from the number in the library less logged exclusions.

2. Introduction

The Peptide Library is the reference set against which Jurassic Meat's reconstructions are inferred and checked. It combines peptide fragments recovered from fossil material by liquid chromatography–tandem mass spectrometry (LC-MS/MS) with protein sequences from extant archosaurs and outgroups. Its purpose is narrow: to provide sequence-level constraints that place the source taxon of each fragment on a reference phylogeny, so that ancestral sequence reconstruction of the muscle and flavour loci we intend to express can be conditioned on that placement. This approach follows the extant phylogenetic bracket as a framework for inference about extinct taxa (Witmer 1995), applied here to protein sequence rather than to soft-tissue anatomy.

The founding peptide result, from a theropod tooth attributed to the Kem Kem Group of southeastern Morocco, was reported in JM-TR-2013-001. Library v1.0 (JM-TR-2014-002) consolidated those fragments with an initial avian and crocodilian reference import, under manual curation. Library v1.4 (JM-TR-2016-003) added fragments from three further specimens with documented provenance. Recovery of protein sequence from Mesozoic material remains contested (Schweitzer et al. 2007; Asara et al. 2007; Buckley et al. 2008; Schweitzer et al. 2009), and our methods have been designed throughout on the assumption that any individual fragment may be exogenous. Evidence that some protein sequence can persist in mineral-bound contexts over deep time (Demarchi et al. 2016), and observations of fibre-like and cell-like structures in Cretaceous dinosaur bone (Bertazzo et al. 2015), inform, but do not settle, that assumption.

Between v1.1 and v1.4 the library was held as one unaligned peptide file per specimen, with alignment to the bracketing coordinate system performed at build time by the automated bracketing module introduced in v1.1 (JM-TR-2014-002, §3.8; JM-TR-2016-003). Bracketing at v1.0 had been manual. This design had three known weaknesses: alignments were not stored and therefore not reproducible across module versions; the same fragment could be aligned differently in successive builds; and reference sequences were imported without version pinning. In May 2017 we decided to migrate to a stored, versioned alignment architecture. This report describes that architecture (Sections 3.2–3.4), the content of v2.0 (Section 4.1–4.3), and a defect identified during migration (Sections 4.4–4.7).

3. Materials and Methods

3.1 Specimens and peptide recovery

Library v2.0 contains fragments from nine fossil specimens. JM-SP-0001 is the founding tooth (JM-TR-2013-001). JM-SP-0011, JM-SP-0014 and JM-SP-0019 were described in JM-TR-2016-003. JM-SP-0031 to JM-SP-0035 (two teeth, two long-bone fragments and one rib fragment) were accessioned between March 2017 and August 2018 under the intake procedure SOP-SP-003, all at provenance Tier A or B (JM-TR-2016-003). No new material from JM-SP-0001 was extracted; its contribution to v2.0 derives from archived raw data only (Section 3.2). Sample preparation followed the palaeoproteomics group's standing procedure: surface removal, demineralisation in weak acid, partitioning of acid-soluble and acid-insoluble fractions, digestion with sequencing-grade trypsin, and analysis on a high-resolution Orbitrap-class instrument. Each extraction batch included at least one procedural blank and one injection blank; blanks were processed identically to specimens. Detailed bench conditions are given in the internal version.

3.2 Re-search of archived raw data

All archived raw LC-MS/MS files from the four v1.4 specimens (n = 212 files) were re-searched under the v2.0 identification workflow: database search with variable deamidation (N/Q), oxidation (M, W), hydroxyproline and non-specific cleavage, against a search space comprising the v2.0 reference set plus a contamination database (human keratins, porcine trypsin, bovine and porcine collagens, common laboratory proteins), followed by de novo sequencing of high-quality unassigned spectra. Peptide-spectrum matches (PSMs) were accepted at 1% false discovery rate (FDR) at peptide level. Feature detection and quantification used MaxQuant-type processing (Cox & Mann 2008). Any peptide matching the contamination database at ≤1 substitution, or present in any blank from the same batch, was excluded and logged.

3.3 Versioned MSA architecture

Library v2.0 stores data as locus partitions: one MSA per locus (currently seven: COL1A1, COL1A2, AMELX, ENAM, AMBN, BGLAP and a pooled partition for other non-collagenous proteins). Columns are keyed to a fixed reference coordinate system derived from the Gallus gallus sequence for each locus, so that a column identifier is stable across releases even when the reference set changes. Each fragment occupies one row, with an explicit gap mask, a per-residue confidence vector carried from the identification step, modification annotations as a separate layer rather than as altered characters, and a pointer to its specimen, extraction batch and raw file.

Partitions are immutable once committed. A library build is defined by a manifest listing the content hash of every partition, the reference-set version, the placement-engine version and the build profile. Two builds with identical manifests are byte-identical. Unaligned legacy peptide files are deprecated as build input (Section 7).

3.4 Fragment alignment and placement

Fragments are aligned to their locus partition by profile alignment against a profile hidden Markov model built from the reference rows, rather than by exact-anchor search (Section 4.5). Fragments shorter than five residues are excluded by rule and logged. Placement uses maximum-likelihood placement on a fixed reference topology, with posterior placement probabilities estimated by likelihood weight ratios and checked against non-parametric bootstrap support (100 replicates) for every fragment of length ≥12. The headline metric reported for each release is the median posterior placement probability on the expected branch (the avian stem) across all theropod-assigned fragments.

3.5 Migration reconciliation

Migration was validated by reconciliation: for every v1.4 fragment, we confirmed (i) presence of a row in the v2.0 MSA, (ii) identity of residue string after removal of modification annotations, and (iii) presence in the bracketing input of a v1.4 reference build. Check (iii) was included to confirm that the v1.4 build could be reproduced. It could not (Section 4.4).

3.6 Retrospective impact assessment

Impact assessment IA-2018-07 was opened on 4 December 2018 under the research organisation's deviation procedure. It covered every reconstruction target produced from a v1.1–v1.4 build, every cell line constructed from those targets, and any material derived from those lines. All 14 targets were re-bracketed using the v1.4 library with the defect corrected and the v1.x engine otherwise unchanged, so that differences could be attributed to the defect alone.

4. Results

4.1 Library content

Library v2.0 comprises 1,284 theropod-assigned fragments: the 389 fragments of v1.4, all retained; 512 additional fragments identified by re-search of archived raw data from the four v1.4 specimens; and 383 fragments from the five new specimens (Table 1). Collagen type I accounts for 79.5% of fragments, consistent with its expected predominance in bone and dentine. Enamel-protein fragments derive only from the five tooth specimens.

Table 1. Library statistics, v1.4 and v2.0.
Metric v1.4 (2016-11) v2.0 (2019-06)
Fossil specimens 4 9
Theropod-assigned fragments 389 1,284
— of which collagen type I (COL1A1/COL1A2) 274 1,021
— of which enamel proteins (AMELX/ENAM/AMBN) 92 182
— of which other non-collagenous proteins 23 81
Unique residues covered 3,050 7,920
Extant reference taxa 19 34
Theropod-diagnostic alignment columns 12 31
Median placement posterior (theropod) 0.47 0.63
Fragments entering bracketing 248 1,267
Fragments in unmapped-fragment log 0 —
Fragments excluded by logged rule not recorded 17

The 17 fragments excluded in v2.0 comprise 11 matching the contamination database at one substitution and six shorter than five residues. All 17 are listed in the build log with reason codes.

4.2 Coverage by specimen and locus

Coverage is uneven (Fig. 1). The founding tooth, JM-SP-0001, remains the best-covered specimen for enamel proteins. Bone specimens contribute collagen almost exclusively. No specimen covers more than half of any locus, and BGLAP coverage is restricted to its short mature peptide.

Figure 1. Fraction of locus residues covered, by specimen (Library v2.0)Fraction of reference-coordinate positions in each locus partition covered by at least one accepted fragment from the specimen. T, tooth; B, bone. Other NCP is a pooled partition.Locus partitionCOL1A1COL1A2AMELXENAMAMBNBGLAPOther NCPJM-SP-0001 (T)JM-SP-0001 (T) × COL1A1: 0.380.38JM-SP-0001 (T) × COL1A2: 0.290.29JM-SP-0001 (T) × AMELX: 0.460.46JM-SP-0001 (T) × ENAM: 0.170.17JM-SP-0001 (T) × AMBN: 0.210.21JM-SP-0001 (T) × BGLAP: 00.00JM-SP-0001 (T) × Other NCP: 0.040.04JM-SP-0011 (T)JM-SP-0011 (T) × COL1A1: 0.220.22JM-SP-0011 (T) × COL1A2: 0.180.18JM-SP-0011 (T) × AMELX: 0.310.31JM-SP-0011 (T) × ENAM: 0.090.09JM-SP-0011 (T) × AMBN: 0.120.12JM-SP-0011 (T) × BGLAP: 00.00JM-SP-0011 (T) × Other NCP: 0.020.02JM-SP-0014 (B)JM-SP-0014 (B) × COL1A1: 0.260.26JM-SP-0014 (B) × COL1A2: 0.210.21JM-SP-0014 (B) × AMELX: 00.00JM-SP-0014 (B) × ENAM: 00.00JM-SP-0014 (B) × AMBN: 00.00JM-SP-0014 (B) × BGLAP: 0.290.29JM-SP-0014 (B) × Other NCP: 0.060.06JM-SP-0019 (T)JM-SP-0019 (T) × COL1A1: 0.140.14JM-SP-0019 (T) × COL1A2: 0.10.10JM-SP-0019 (T) × AMELX: 0.190.19JM-SP-0019 (T) × ENAM: 0.050.05JM-SP-0019 (T) × AMBN: 0.060.06JM-SP-0019 (T) × BGLAP: 00.00JM-SP-0019 (T) × Other NCP: 0.010.01JM-SP-0031 (T)JM-SP-0031 (T) × COL1A1: 0.240.24JM-SP-0031 (T) × COL1A2: 0.20.20JM-SP-0031 (T) × AMELX: 0.350.35JM-SP-0031 (T) × ENAM: 0.130.13JM-SP-0031 (T) × AMBN: 0.150.15JM-SP-0031 (T) × BGLAP: 00.00JM-SP-0031 (T) × Other NCP: 0.030.03JM-SP-0032 (T)JM-SP-0032 (T) × COL1A1: 0.170.17JM-SP-0032 (T) × COL1A2: 0.110.11JM-SP-0032 (T) × AMELX: 0.220.22JM-SP-0032 (T) × ENAM: 0.060.06JM-SP-0032 (T) × AMBN: 0.070.07JM-SP-0032 (T) × BGLAP: 00.00JM-SP-0032 (T) × Other NCP: 00.00JM-SP-0033 (B)JM-SP-0033 (B) × COL1A1: 0.420.42JM-SP-0033 (B) × COL1A2: 0.340.34JM-SP-0033 (B) × AMELX: 00.00JM-SP-0033 (B) × ENAM: 00.00JM-SP-0033 (B) × AMBN: 00.00JM-SP-0033 (B) × BGLAP: 0.410.41JM-SP-0033 (B) × Other NCP: 0.080.08JM-SP-0034 (B)JM-SP-0034 (B) × COL1A1: 0.280.28JM-SP-0034 (B) × COL1A2: 0.210.21JM-SP-0034 (B) × AMELX: 00.00JM-SP-0034 (B) × ENAM: 00.00JM-SP-0034 (B) × AMBN: 00.00JM-SP-0034 (B) × BGLAP: 00.00JM-SP-0034 (B) × Other NCP: 0.050.05JM-SP-0035 (B)JM-SP-0035 (B) × COL1A1: 0.330.33JM-SP-0035 (B) × COL1A2: 0.270.27JM-SP-0035 (B) × AMELX: 00.00JM-SP-0035 (B) × ENAM: 00.00JM-SP-0035 (B) × AMBN: 00.00JM-SP-0035 (B) × BGLAP: 0.330.33JM-SP-0035 (B) × Other NCP: 0.070.07Figure 1. Fraction of locus residues covered, by specimen (Library v2.0)Fraction of reference-coordinate positions in each locus partition covered by at least one accepted fragment from the specimen. T, tooth; B, bone. Other NCP is a pooled partition.Locus partitionCOL1A1COL1A2AMELXENAMAMBNBGLAPOther NCPJM-SP-0001 (T)JM-SP-0001 (T) × COL1A1: 0.38JM-SP-0001 (T) × COL1A2: 0.29JM-SP-0001 (T) × AMELX: 0.46JM-SP-0001 (T) × ENAM: 0.17JM-SP-0001 (T) × AMBN: 0.21JM-SP-0001 (T) × BGLAP: 0JM-SP-0001 (T) × Other NCP: 0.04JM-SP-0011 (T)JM-SP-0011 (T) × COL1A1: 0.22JM-SP-0011 (T) × COL1A2: 0.18JM-SP-0011 (T) × AMELX: 0.31JM-SP-0011 (T) × ENAM: 0.09JM-SP-0011 (T) × AMBN: 0.12JM-SP-0011 (T) × BGLAP: 0JM-SP-0011 (T) × Other NCP: 0.02JM-SP-0014 (B)JM-SP-0014 (B) × COL1A1: 0.26JM-SP-0014 (B) × COL1A2: 0.21JM-SP-0014 (B) × AMELX: 0JM-SP-0014 (B) × ENAM: 0JM-SP-0014 (B) × AMBN: 0JM-SP-0014 (B) × BGLAP: 0.29JM-SP-0014 (B) × Other NCP: 0.06JM-SP-0019 (T)JM-SP-0019 (T) × COL1A1: 0.14JM-SP-0019 (T) × COL1A2: 0.1JM-SP-0019 (T) × AMELX: 0.19JM-SP-0019 (T) × ENAM: 0.05JM-SP-0019 (T) × AMBN: 0.06JM-SP-0019 (T) × BGLAP: 0JM-SP-0019 (T) × Other NCP: 0.01JM-SP-0031 (T)JM-SP-0031 (T) × COL1A1: 0.24JM-SP-0031 (T) × COL1A2: 0.2JM-SP-0031 (T) × AMELX: 0.35JM-SP-0031 (T) × ENAM: 0.13JM-SP-0031 (T) × AMBN: 0.15JM-SP-0031 (T) × BGLAP: 0JM-SP-0031 (T) × Other NCP: 0.03JM-SP-0032 (T)JM-SP-0032 (T) × COL1A1: 0.17JM-SP-0032 (T) × COL1A2: 0.11JM-SP-0032 (T) × AMELX: 0.22JM-SP-0032 (T) × ENAM: 0.06JM-SP-0032 (T) × AMBN: 0.07JM-SP-0032 (T) × BGLAP: 0JM-SP-0032 (T) × Other NCP: 0JM-SP-0033 (B)JM-SP-0033 (B) × COL1A1: 0.42JM-SP-0033 (B) × COL1A2: 0.34JM-SP-0033 (B) × AMELX: 0JM-SP-0033 (B) × ENAM: 0JM-SP-0033 (B) × AMBN: 0JM-SP-0033 (B) × BGLAP: 0.41JM-SP-0033 (B) × Other NCP: 0.08JM-SP-0034 (B)JM-SP-0034 (B) × COL1A1: 0.28JM-SP-0034 (B) × COL1A2: 0.21JM-SP-0034 (B) × AMELX: 0JM-SP-0034 (B) × ENAM: 0JM-SP-0034 (B) × AMBN: 0JM-SP-0034 (B) × BGLAP: 0JM-SP-0034 (B) × Other NCP: 0.05JM-SP-0035 (B)JM-SP-0035 (B) × COL1A1: 0.33JM-SP-0035 (B) × COL1A2: 0.27JM-SP-0035 (B) × AMELX: 0JM-SP-0035 (B) × ENAM: 0JM-SP-0035 (B) × AMBN: 0JM-SP-0035 (B) × BGLAP: 0.33JM-SP-0035 (B) × Other NCP: 0.07
Figure 1.Fraction of locus residues covered, by specimen (Library v2.0). Fraction of reference-coordinate positions in each locus partition covered by at least one accepted fragment from the specimen. T, tooth; B, bone. Other NCP is a pooled partition.Source: JM Shenzhen Research Center, internal dataset PL-2019-06.
Source data
SpecimenCOL1A1COL1A2AMELXENAMAMBNBGLAPOther NCP
JM-SP-0001 (T)0.380.290.460.170.2100.04
JM-SP-0011 (T)0.220.180.310.090.1200.02
JM-SP-0014 (B)0.260.210000.290.06
JM-SP-0019 (T)0.140.10.190.050.0600.01
JM-SP-0031 (T)0.240.20.350.130.1500.03
JM-SP-0032 (T)0.170.110.220.060.0700
JM-SP-0033 (B)0.420.340000.410.08
JM-SP-0034 (B)0.280.2100000.05
JM-SP-0035 (B)0.330.270000.330.07

4.3 Placement performance

Median posterior placement probability for theropod-assigned fragments rose from 0.47 (v1.4) to 0.63 (v2.0). Approximately two-thirds of the improvement is attributable to the expanded reference set (19 to 34 taxa, including three further palaeognaths, for five in total, and four additional crocodilians), and the remainder to profile alignment and the increase in fragment count. Posterior probability depends strongly on fragment length (Fig. 2). Fragments of six and seven residues remain poorly placed in both versions, with wide interquartile ranges. For comparability, the v1.4 curve in Fig. 2 is computed by re-scoring the v1.4 library with the v2.0 engine; it therefore differs from the curve reported in JM-TR-2016-003, which used the v1.x placement pipeline.

Figure 2. Placement confidence by fragment lengthMedian posterior placement probability (line) and interquartile range (band) for theropod-assigned fragments: v1.4 library re-scored with the v2.0 engine, and the v2.0 library. n = 389 and 1,284 fragments respectively.Posterior placement probability0.000.200.400.600.801.00681012141618202224Fragment length (residues)v1.x anchor seed length (8)v1.4 (re-scored)v2.0Fragment length (residues): 6 v1.4 (re-scored): 0.19 v2.0: 0.24 (0.07–0.49)Fragment length (residues): 7 v1.4 (re-scored): 0.25 v2.0: 0.33 (0.12–0.55)Fragment length (residues): 8 v1.4 (re-scored): 0.34 v2.0: 0.46 (0.29–0.62)Fragment length (residues): 9 v1.4 (re-scored): 0.42 v2.0: 0.58 (0.41–0.71)Fragment length (residues): 10 v1.4 (re-scored): 0.49 v2.0: 0.66 (0.52–0.77)Fragment length (residues): 12 v1.4 (re-scored): 0.58 v2.0: 0.74 (0.62–0.83)Fragment length (residues): 14 v1.4 (re-scored): 0.65 v2.0: 0.80 (0.70–0.87)Fragment length (residues): 16 v1.4 (re-scored): 0.70 v2.0: 0.84 (0.76–0.90)Fragment length (residues): 20 v1.4 (re-scored): 0.77 v2.0: 0.88 (0.81–0.93)Fragment length (residues): 24 v1.4 (re-scored): 0.81 v2.0: 0.91 (0.86–0.95)Figure 2. Placement confidence by fragment lengthMedian posterior placement probability (line) and interquartile range (band) for theropod-assigned fragments: v1.4 library re-scored with the v2.0 engine, and the v2.0 library. n = 389 and 1,284 fragments respectively.Posterior placement probability0.000.200.400.600.801.00101520Fragment length (residues)v1.x anchor seedlength (8)Fragment length (residues): 6 v1.4 (re-scored): 0.19 v2.0: 0.24 (0.07–0.49)Fragment length (residues): 7 v1.4 (re-scored): 0.25 v2.0: 0.33 (0.12–0.55)Fragment length (residues): 8 v1.4 (re-scored): 0.34 v2.0: 0.46 (0.29–0.62)Fragment length (residues): 9 v1.4 (re-scored): 0.42 v2.0: 0.58 (0.41–0.71)Fragment length (residues): 10 v1.4 (re-scored): 0.49 v2.0: 0.66 (0.52–0.77)Fragment length (residues): 12 v1.4 (re-scored): 0.58 v2.0: 0.74 (0.62–0.83)Fragment length (residues): 14 v1.4 (re-scored): 0.65 v2.0: 0.80 (0.70–0.87)Fragment length (residues): 16 v1.4 (re-scored): 0.70 v2.0: 0.84 (0.76–0.90)Fragment length (residues): 20 v1.4 (re-scored): 0.77 v2.0: 0.88 (0.81–0.93)Fragment length (residues): 24 v1.4 (re-scored): 0.81 v2.0: 0.91 (0.86–0.95)
Figure 2.Placement confidence by fragment length. Median posterior placement probability (line) and interquartile range (band) for theropod-assigned fragments: v1.4 library re-scored with the v2.0 engine, and the v2.0 library. n = 389 and 1,284 fragments respectively.Source: JM Shenzhen Research Center, internal dataset PL-2019-06.
Source data
Fragment length (residues)v1.4 (re-scored)v2.0v2.0 interval
60.190.240.07–0.49
70.250.330.12–0.55
80.340.460.29–0.62
90.420.580.41–0.71
100.490.660.52–0.77
120.580.740.62–0.83
140.650.80.7–0.87
160.70.840.76–0.9
200.770.880.81–0.93
240.810.910.86–0.95

The consensus placement of theropod-assigned fragments on the reference topology is shown in Fig. 3. The placement is on the avian stem, as expected for a non-avian theropod, with a posterior of 0.63 on that branch and the bulk of the residual mass distributed to the crown-bird root (0.21) and the archosaur root (0.11).

Figure 3. Consensus placement of theropod-assigned fragments, Library v2.0Simplified reference topology (12 of 34 extant reference taxa shown). JM_theropod denotes the consensus placement of all theropod-assigned fragments; p is the median posterior on the avian stem. Outgroup: Chrysemys.JM theropodp = 0.63JM theropod — placement posterior 0.63StruthioStruthioDromaiusDromaiusTinamusTinamusGallusGallusMeleagrisMeleagrisAnasAnasTaeniopygiaTaeniopygiaAlligatorAlligatorCaimanCaimanCrocodylusCrocodylusGavialisGavialisChrysemysChrysemysFigure 3. Consensus placement of theropod-assigned fragments, Library v2.0Simplified reference topology (12 of 34 extant reference taxa shown). JM_theropod denotes the consensus placement of all theropod-assigned fragments; p is the median posterior on the avian stem. Outgroup: Chrysemys.JM theropodp = 0.63JM theropod — placement posterior 0.63StruthioStruthioDromaiusDromaiusTinamusTinamusGallusGallusMeleagrisMeleagrisAnasAnasTaeniopygiaTaeniopygiaAlligatorAlligatorCaimanCaimanCrocodylusCrocodylusGavialisGavialisChrysemysChrysemys
Figure 3.Consensus placement of theropod-assigned fragments, Library v2.0. Simplified reference topology (12 of 34 extant reference taxa shown). JM_theropod denotes the consensus placement of all theropod-assigned fragments; p is the median posterior on the avian stem. Outgroup: Chrysemys.Source: JM Shenzhen Research Center, internal dataset PL-2019-06.
Source data
TaxonHighlightedPlacement posterior
JM theropodyes0.63
Struthio
Dromaius
Tinamus
Gallus
Meleagris
Anas
Taeniopygia
Alligator
Caiman
Crocodylus
Gavialis
Chrysemys

(((JM_theropod,((Struthio,(Dromaius,Tinamus)),(((Gallus,Meleagris),Anas),Taeniopygia))),((Alligator,Caiman),(Crocodylus,Gavialis))),Chrysemys);

4.4 Detection of the fragment-loss defect

Reconciliation check (iii) (Section 3.5) was run on 21 November 2018 against the archived v1.4 reference build. Of the 389 fragments in v1.4, 248 appeared in the bracketing input. The remaining 141 had rows in the migrated MSA, matched their residue strings, and were absent from bracketing. None appeared in the unmapped-fragment log, which records fragments the bracketing module cannot map. That log was empty for every build from v1.1 to v1.4, and its emptiness had been taken, where it was consulted, as confirmation that every fragment had mapped. Review of the log was not part of the release procedure (JM-TR-2014-002; JM-TR-2016-003). The v1.4 build log reported "constraints applied: 248". Figures of 154, 171 and 194 had been reported for v1.1 to v1.3 (JM-TR-2016-003, Section 6). None had been compared with the library fragment count.

The same check, applied to the archived builds of the maintenance releases v1.1 to v1.3, which added 87 fragments from further extraction campaigns on JM-SP-0001 (JM-TR-2016-003, Table S1), showed the defect in all of them (Table 2; Fig. 4). The v1.0 build was unaffected: bracketing at v1.0 was manual and predates the module described in Section 4.5, and the curators confirmed at release that all 214 fragments had been entered (JM-TR-2014-002).

Table 2. Fragments silently omitted from bracketing, by build and cause.
Build Date Fragments in library Entered bracketing Omitted Length <8 Annotated deamidation I/L ambiguity code Multiple causes
v1.0 2014-03 214 214 0 — — — —
v1.1 2014-09 238 154 84 28 30 15 11
v1.2 2015-04 263 171 92 32 33 15 12
v1.3 2015-12 301 194 107 41 36 17 13
v1.4 2016-11 389 248 141 58 47 21 15

The 141 fragments omitted in v1.4 comprise the 107 JM-SP-0001 fragments omitted at v1.3 and 34 of the 88 fragments from the three new specimens, which are shorter on average (JM-TR-2016-003). All 15 multiple-cause fragments carried annotated deamidation. In total, 62 of the 104 v1.4 fragments carrying at least one annotated deamidation (59.6%) were omitted, compared with 79 of the 285 fragments without (27.7%). The defect therefore preferentially excluded fragments bearing the modification most often cited in support of endogeneity.

Figure 4. Fragments entering and omitted from bracketing, builds v1.0–v1.4Stacked counts per build. v1.1–v1.3 were maintenance releases adding fragments from further extraction of JM-SP-0001. No omission appeared in the unmapped-fragment log of any build; all were identified by reconciliation in November 2018.Fragments0100200300400v1.0v1.0 Entered bracketing: 214 Total: 214214v1.1v1.1 Entered bracketing: 154 Total: 238v1.1 Silently omitted: 84 Total: 238238v1.2v1.2 Entered bracketing: 171 Total: 263v1.2 Silently omitted: 92 Total: 263263v1.3v1.3 Entered bracketing: 194 Total: 301v1.3 Silently omitted: 107 Total: 301301v1.4v1.4 Entered bracketing: 248 Total: 389v1.4 Silently omitted: 141 Total: 389389BuildFigure 4. Fragments entering and omitted from bracketing, builds v1.0–v1.4Stacked counts per build. v1.1–v1.3 were maintenance releases adding fragments from further extraction of JM-SP-0001. No omission appeared in the unmapped-fragment log of any build; all were identified by reconciliation in November 2018.Fragments0100200300400v1.0v1.0 Entered bracketing: 214 Total: 214214v1.1v1.1 Entered bracketing: 154 Total: 238v1.1 Silently omitted: 84 Total: 238238v1.2v1.2 Entered bracketing: 171 Total: 263v1.2 Silently omitted: 92 Total: 263263v1.3v1.3 Entered bracketing: 194 Total: 301v1.3 Silently omitted: 107 Total: 301301v1.4v1.4 Entered bracketing: 248 Total: 389v1.4 Silently omitted: 141 Total: 389389Build
Figure 4.Fragments entering and omitted from bracketing, builds v1.0–v1.4. Stacked counts per build. v1.1–v1.3 were maintenance releases adding fragments from further extraction of JM-SP-0001. No omission appeared in the unmapped-fragment log of any build; all were identified by reconciliation in November 2018.Source: Archived build logs and reconciliation output, IA-2018-07.
Source data
BuildEntered bracketingSilently omitted
v1.02140
v1.115484
v1.217192
v1.3194107
v1.4248141

The length distribution of the omitted fragments is shown in Fig. 5. Fragments of six and seven residues account for 58 of 141; the remainder are of ordinary length and were omitted for annotation reasons alone.

Figure 5. Length distribution of fragments omitted from bracketing (v1.4 build)n = 141. Fragments ≥8 residues were omitted because an annotated modification or ambiguity code prevented an exact anchor match.Fragments05101520253035678101214162024326–7: 277–8: 318–10: 2410–12: 2112–14: 1514–16: 1016–20: 820–24: 424–32: 1Fragment length (residues)Figure 5. Length distribution of fragments omitted from bracketing (v1.4 build)n = 141. Fragments ≥8 residues were omitted because an annotated modification or ambiguity code prevented an exact anchor match.Fragments01020304010152025306–7: 277–8: 318–10: 2410–12: 2112–14: 1514–16: 1016–20: 820–24: 424–32: 1Fragment length (residues)
Figure 5.Length distribution of fragments omitted from bracketing (v1.4 build). n = 141. Fragments ≥8 residues were omitted because an annotated modification or ambiguity code prevented an exact anchor match.Source: Reconciliation output, IA-2018-07.
Source data
Fragment length (residues)Fragments
6–727
7–831
8–1024
10–1221
12–1415
14–1610
16–208
20–244
24–321

4.5 Root cause

On 19 August 2014 the automated bracketing module was completed for v1.1, in preparation for fragments from additional specimens, which were expected to arrive without the hand-resolved annotations of the v1.0 entries. The module added a normaliser for residue strings and, ahead of its mapping step, a fast pre-filter that located each fragment in its reference locus by an exact match of an eight-residue seed. Three conditions caused the search to return an empty anchor set: a fragment shorter than the seed; a fragment whose residue string contained a lowercase modification annotation (the v1.x convention for deamidated N and Q), which the normaliser converted to the unknown-residue character; and a fragment containing the I/L ambiguity code. In each case the calling function interpreted an empty anchor set as "fragment contributes no constraint", which is the correct interpretation for a fragment that is uninformative, and skipped the fragment without raising an error. Because the skip occurred in the pre-filter, the fragment never reached the mapping step, which is the only code path that writes to the unmapped-fragment log.

The design assumed that the seed search would fail only for fragments that could not be aligned at all. This assumption was documented in a code comment and was not tested. Unit tests for the module used a synthetic fragment set of length ≥10 with no annotations.

4.6 Impact on reconstruction targets

All 14 reconstruction targets produced between September 2014 and November 2018 were re-bracketed with the defect corrected (Section 3.6). Restoring the 141 fragments altered the topology-conditional branch lengths on the avian stem and therefore the ancestral-state posteriors at weakly constrained sites. Inferred residues changed at 37 positions in nine targets (Table 3). Of the 37 changes, 31 occurred at sites whose original posterior was below 0.6; six occurred at sites with original posterior above 0.9. The number of changes per target scales with the number of weakly constrained sites in that target (Fig. 6).

Table 3. Effect of re-bracketing on reconstruction targets built from v1.1–v1.4.
Target Gene product Sites with posterior <0.9 Residues changed Used in research line
MB Myoglobin 11 3 DN-2016-03
ACTA1 Skeletal α-actin 2 0 DN-2016-03
MYH1 Myosin heavy chain, fast IIx 38 9 DN-2016-03, DN-2017-01
MYH7 Myosin heavy chain, slow 21 4 —
TNNT3 Troponin T, fast 9 2 DN-2017-01
TNNI2 Troponin I, fast 3 0 —
MYL1 Myosin light chain 1 6 1 DN-2017-01
MYOD1 Myogenic differentiation 1 4 0 DN-2016-03
MYF5 Myogenic factor 5 5 0 —
MSTN Myostatin 17 5 DN-2017-01
CKM Creatine kinase, muscle 8 2 —
FABP4 Fatty-acid-binding protein 4 3 0 —
LPL Lipoprotein lipase 26 7 —
PPARG PPAR-γ 15 4 —
Total 168 37

Figure 6. Residue changes on re-bracketing versus weakly constrained sites, per targetEach point is one of 14 reconstruction targets built from v1.1–v1.4. Linear fit shown (slope 0.27 changes per weakly constrained site; R² = 0.96).Residues changed after correction02468100510152025303540Sites with ancestral-state posterior < 0.9Reconstruction targets linear fit: y = 0.268x − 0.576, R² = 0.96MB Reconstruction targets Sites with ancestral-state posterior < 0.9: 11 Residues changed after correction: 3MBACTA1 Reconstruction targets Sites with ancestral-state posterior < 0.9: 2 Residues changed after correction: 0MYH1 Reconstruction targets Sites with ancestral-state posterior < 0.9: 38 Residues changed after correction: 9MYH1MYH7 Reconstruction targets Sites with ancestral-state posterior < 0.9: 21 Residues changed after correction: 4MYH7TNNT3 Reconstruction targets Sites with ancestral-state posterior < 0.9: 9 Residues changed after correction: 2TNNT3TNNI2 Reconstruction targets Sites with ancestral-state posterior < 0.9: 3 Residues changed after correction: 0MYL1 Reconstruction targets Sites with ancestral-state posterior < 0.9: 6 Residues changed after correction: 1MYL1MYOD1 Reconstruction targets Sites with ancestral-state posterior < 0.9: 4 Residues changed after correction: 0MYF5 Reconstruction targets Sites with ancestral-state posterior < 0.9: 5 Residues changed after correction: 0MYF5MSTN Reconstruction targets Sites with ancestral-state posterior < 0.9: 17 Residues changed after correction: 5MSTNCKM Reconstruction targets Sites with ancestral-state posterior < 0.9: 8 Residues changed after correction: 2FABP4 Reconstruction targets Sites with ancestral-state posterior < 0.9: 3 Residues changed after correction: 0LPL Reconstruction targets Sites with ancestral-state posterior < 0.9: 26 Residues changed after correction: 7LPLPPARG Reconstruction targets Sites with ancestral-state posterior < 0.9: 15 Residues changed after correction: 4PPARGFigure 6. Residue changes on re-bracketing versus weakly constrained sites, per targetEach point is one of 14 reconstruction targets built from v1.1–v1.4. Linear fit shown (slope 0.27 changes per weakly constrained site; R² = 0.96).Residues changed after correction0246810010203040Sites with ancestral-state posterior < 0.9Reconstruction targets linear fit: y = 0.268x − 0.576, R² = 0.96MB Reconstruction targets Sites with ancestral-state posterior < 0.9: 11 Residues changed after correction: 3MBACTA1 Reconstruction targets Sites with ancestral-state posterior < 0.9: 2 Residues changed after correction: 0MYH1 Reconstruction targets Sites with ancestral-state posterior < 0.9: 38 Residues changed after correction: 9MYH1MYH7 Reconstruction targets Sites with ancestral-state posterior < 0.9: 21 Residues changed after correction: 4MYH7TNNT3 Reconstruction targets Sites with ancestral-state posterior < 0.9: 9 Residues changed after correction: 2TNNT3TNNI2 Reconstruction targets Sites with ancestral-state posterior < 0.9: 3 Residues changed after correction: 0MYL1 Reconstruction targets Sites with ancestral-state posterior < 0.9: 6 Residues changed after correction: 1MYL1MYOD1 Reconstruction targets Sites with ancestral-state posterior < 0.9: 4 Residues changed after correction: 0MYF5 Reconstruction targets Sites with ancestral-state posterior < 0.9: 5 Residues changed after correction: 0MYF5MSTN Reconstruction targets Sites with ancestral-state posterior < 0.9: 17 Residues changed after correction: 5MSTNCKM Reconstruction targets Sites with ancestral-state posterior < 0.9: 8 Residues changed after correction: 2FABP4 Reconstruction targets Sites with ancestral-state posterior < 0.9: 3 Residues changed after correction: 0LPL Reconstruction targets Sites with ancestral-state posterior < 0.9: 26 Residues changed after correction: 7LPLPPARG Reconstruction targets Sites with ancestral-state posterior < 0.9: 15 Residues changed after correction: 4PPARG
Figure 6.Residue changes on re-bracketing versus weakly constrained sites, per target. Each point is one of 14 reconstruction targets built from v1.1–v1.4. Linear fit shown (slope 0.27 changes per weakly constrained site; R² = 0.96).Source: Re-bracketing output, IA-2018-07.
Source data
LabelSites with ancestral-state posterior < 0.9Residues changed after correction
MB113
ACTA120
MYH1389
MYH7214
TNNT392
TNNI230
MYL161
MYOD140
MYF550
MSTN175
CKM82
FABP430
LPL267
PPARG154

Two research cell lines had been constructed from affected targets. DN-2016-03 incorporated targets MB, ACTA1, MYH1 and MYOD1; DN-2017-01 incorporated MYH1, TNNT3, MYL1 and MSTN. Both lines were retired in 2018 for reasons unrelated to the defect (proliferation rate below specification in extended passage). No material from either line was used outside the research organisation. Construction of DN-2019-04 was initiated in April 2019 against release candidate v2.0-rc2, whose targets are identical to those of v2.0 final, and is unaffected.

The v1.4 placement statistics reported in JM-TR-2016-003, including the median posterior of 0.47, were computed by the placement pipeline, which does not use the bracketing module, and are unaffected.

4.7 Impact assessment conclusion

IA-2018-07 was closed on 12 March 2019 with the following findings: (a) the defect affected all builds from v1.1 to v1.4; (b) 37 inferred residues in nine reconstruction targets were affected; (c) two research cell lines had incorporated affected targets and had been retired for unrelated reasons; (d) no material derived from any affected build had been released. Conclusion: no impact on released material. The assessment was presented to the Ethics & Bioresponsibility Council on 19 March 2019: reviewed; no concerns raised. The full incident timeline is given in Table 4.

Table 4. Incident timeline.
Date Event
2014-03-17 Library v1.0 released; bracketing manual; all 214 fragments entered by hand
2014-08-19 Automated bracketing module completed; normaliser and exact eight-residue seed pre-filter
2014-09-16 First automated build (v1.1); 84 of 238 fragments omitted; unmapped-fragment log empty
2016-11-08 Library v1.4 released; 141 of 389 fragments omitted; unmapped-fragment log empty, not reviewed
2017-05-09 Decision to migrate to versioned MSA architecture
2018-11-21 Reconciliation check (iii) fails: 248 of 389 fragments in v1.4 bracketing input
2018-11-23 Root cause identified (Section 4.5)
2018-12-04 Impact assessment IA-2018-07 opened
2019-01-08 → 2019-02-22 Re-bracketing of 14 reconstruction targets
2019-03-12 IA-2018-07 closed: no impact on released material
2019-03-19 Ethics & Bioresponsibility Council: reviewed; no concerns raised
2019-04-15 DN-2019-04 construction initiated against v2.0-rc2
2019-06-04 Library v2.0 released

Elapsed time from introduction of the defect to detection: 51 months.

5. Discussion

The v2.0 architecture addresses the reproducibility problems that motivated it. Every build is now identified by a manifest hash and can be regenerated exactly; alignment is stored rather than recomputed; modifications are an annotation layer rather than altered characters; and the reference set is pinned. The increase in median placement posterior from 0.47 to 0.63 is the largest single-release improvement to date and is driven principally by reference-set expansion, in particular the addition of paleognath sequences, which shorten the branch to which theropod-assigned fragments attach.

The fragment-loss defect is the more consequential finding of this work. It persisted for 51 months because its only observable symptom, a "constraints applied" count lower than the fragment count, was present in every build log and was never compared with anything, while the log designed to record unmapped fragments remained empty. An empty log was read as evidence that the scripts were working. The defect affected fragments non-randomly: short fragments, which carry little information, and deamidated fragments, which carry the strongest available signal of age. For approximately four years, therefore, our bracketing builds were conditioned on a subset of the data that was shorter of evidence of endogeneity than the library as a whole. We note that this did not alter the topology of placement in any build, and that the placement statistics reported for v1.4 were unaffected (Section 4.6).

The effect on reconstruction targets was modest in magnitude and concentrated where it would be expected, at sites already identified as weakly constrained. The six changes at high-posterior sites are documented in the internal version. We regard the retirement of DN-2016-03 and DN-2017-01 before detection as fortunate rather than as a control.

The corrective actions (Table 5) follow a single principle: the library must account for every fragment in every build. A fragment may be excluded, but it may not be excluded silently.

Table 5. Corrective and preventive actions.
ID Action Owner Status at release
CA-1 Replace exact-seed anchor search with profile-HMM alignment Computational Genomics Complete (v2.0)
CA-2 Empty alignment result raises an error and is never interpreted as "no constraint" Computational Genomics Complete (v2.0)
CA-3 Build fails if fragments entering bracketing ≠ library fragments − logged exclusions Computational Genomics Complete (v2.0)
CA-4 Every exclusion logged with a reason code and listed in the release manifest Computational Genomics Complete (v2.0)
CA-5 Re-bracket all 14 v1.x reconstruction targets; supersede affected versions Genome Reconstruction Complete
CA-6 Convert the 41 v1.x override flags, and every manual edit in legacy files that departs from automated output, to structured override records (SOP-PL-003 rev. D); rationale field recommended Palaeoproteomics Complete (v2.0)
CA-7 Unit tests for alignment to include short, annotated and ambiguous fragments Computational Genomics Complete
CA-8 Periodic review of build logs Computational Genomics Scheduled

6. Limitations and known issues

  • Fragments of six and seven residues are now included in bracketing but are poorly placed (median posterior 0.24 and 0.33; Fig. 2). Their contribution to the joint likelihood is small, and their variance is high. A dedicated treatment is planned for a future release.
  • Modification annotations are now carried as a separate layer, but placement scoring does not yet model diagenetic modification; a deamidated residue is scored as the observed residue. Diagenetic modelling is planned for a future release.
  • At migration, the 41 override flags carried from v1.x (37 applied at v1.0 and four at v1.2; JM-TR-2014-002; JM-TR-2016-003) were converted to structured override records. A further 1,166 manual edits found in the legacy peptide files, which departed from automated output but carried no flag, were converted to override records so that v2.0 builds would reproduce v1.4 behaviour (CA-6). All 1,207 legacy records are attributed to the migration service account jm-curate, as individual attribution was not recorded at the time. Overrides take precedence over automated behaviour and do not expire.
  • Re-search of archived raw data used a contamination database current to 2019. Fragments accepted in v1.0 under the 2014 database were re-screened; none was excluded.
  • JM-SP-0001 remains provenance Tier C (JM-TR-2016-003). Its record is incomplete for the period before receipt at the laboratory in 2011.
  • Build filters for clade-targeted builds are specified in the build profile by the user and are not validated against the target clade.
  • Six inferred residues at sites with original posterior above 0.9 changed on correction (Section 4.6). These are documented in the internal version and were not investigated further, as the affected research lines had been retired.

7. Changes in Library v2.0

  • Breaking change: migrated to a versioned multiple-sequence alignment format.
  • Deprecated support for unaligned legacy peptide files.
  • Fixed an issue where fragmentary sequences were silently dropped during bracketing.

The versioned format is described in Section 3.3; legacy file deprecation in Section 3.3; the fixed issue in Sections 4.4–4.7.

8. Data availability

Library contents are confidential. Library access is restricted to authorized research staff and is not available to third parties. Aggregate statistics in this report are derived from internal dataset PL-2019-06 and build manifest v2.0 (content hash recorded in the internal version). Questions about this report may be directed to regulatory@jurassicmeat.com.

9. Acknowledgements

We thank the members of the Palaeoproteomics and Computational Genomics groups who ran the reconciliation checks, and Marcus Oyelaran for retrieval of archived build media from controlled storage. The impact assessment was reviewed by the Ethics & Bioresponsibility Council: reviewed; no concerns raised. This work was funded internally. All authors are employees of Jurassic Meat and may hold equity.

10. References

  1. Asara JM, et al. (2007). Protein sequences from mastodon and Tyrannosaurus rex revealed by mass spectrometry. Science 316:280–285.
  2. Bertazzo S, et al. (2015). Fibres and cellular structures preserved in 75-million-year-old dinosaur specimens. Nature Communications 6:7352.
  3. Buckley M, et al. (2008). Comment on "Protein sequences from mastodon and Tyrannosaurus rex revealed by mass spectrometry." Science 319:33.
  4. Cox J, Mann M (2008). MaxQuant enables high peptide identification rates, individualized p.p.b.-range mass accuracies and proteome-wide protein quantification. Nature Biotechnology 26:1367–1372.
  5. Demarchi B, et al. (2016). Protein sequences bound to mineral surfaces persist into deep time. eLife 5:e17092.
  6. Schweitzer MH, et al. (2007). Analyses of soft tissue from Tyrannosaurus rex suggest the presence of protein. Science 316:277–280.
  7. Schweitzer MH, et al. (2009). Biomolecular characterization and protein sequences of the Campanian hadrosaur B. canadensis. Science 324:626–631.
  8. Witmer LM (1995). The extant phylogenetic bracket and the importance of reconstructing soft tissues in fossils. In: Thomason JJ (ed.), Functional Morphology in Vertebrate Paleontology. Cambridge University Press, 19–33.
  9. JM-TR-2013-001. Endogenous peptide fragments recovered from a Kem Kem Group theropod tooth. Jurassic Meat internal technical report.
  10. JM-TR-2014-002. Library v1.0: first structured release. Jurassic Meat internal technical report.
  11. JM-TR-2016-003. Library v1.4: fragments from additional specimens with documented provenance. Jurassic Meat internal technical report.

Supplementary Table S1

Table S1. Extant reference taxa added in v2.0 (15 of 34; the 19 carried from v1.4 are listed in JM-TR-2016-003, Table S2).
Group Taxa added
Palaeognaths (3) Dromaius novaehollandiae, Rhea americana, Apteryx australis
Galloanserae (2) Coturnix japonica, Anser cygnoides
Neoaves (6) Aquila chrysaetos, Picoides pubescens, Pygoscelis adeliae, Cuculus canorus, Tyto alba, Nipponia nippon
Crocodilians (4) Alligator sinensis, Crocodylus niloticus, Crocodylus siamensis, Melanosuchus niger

With these additions the v2.0 reference set comprises five palaeognaths (Struthio camelus and Tinamus guttatus carried from v1.4), five Galloanserae, 13 Neoaves, eight crocodilians and three non-archosaur outgroups.